Overall Strategy

The main idea was to make a model as simple as possible without the complication of preprocessing or postprocessing. For training the model, I used only one image of a ball which I found using a simple OpenCV function. After this, I pasted the same ball onto background photos I took from Kaggle. I divided the ball into 2 classes: magic-ball and a fixed color circle/ball. There are other wide-ranging possibilities to perform the task, but I chose this way because, in my understanding, it is the fastest (as you can see, it takes me an average of 70 ms per image) and requires minimum time for pre- and post-processing.

• Step 1 - I chose a simple image and using HoughCircles (OpenCV) found the magic ball.
• Step 2 - Using copy-paste augmentation, I created a new dataset.
• Step 3 - I trained a YOLOv8-nano detector to find 2 classes (0: magic ball, 1: regular ball).
• Step 4 - For inference, I used scale TTA and NMS for ensemble.
• Step 5 - For postprocessing, I converted the bounding boxes to circles.

Datastes and requirements

• kaggle datasets download -d balraj98/stanford-background-dataset & unzip stanford-background-dataset -d stanford-background-dataset
• pip install -U ultralytics
• pip install torch==1.8.1+cu111 torchvision==0.9.1+cu111 torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
• pip install numpy opencv-python
• pip install ensemble-boxes

from IPython.display import clear_output
import torch
import cv2
import numpy as np
import albumentations as A
import cv2
from matplotlib import pyplot as plt
from PIL import Image, ImageDraw
import sys
import glob
from tqdm import tqdm
from ultralytics import YOLO
import time
from ensemble_boxes import *
import os

class config:
    magic_ball_for_example = 'Balls/MVC-001F.JPG'
    random_paths = glob.glob('stanford-background-dataset/images/*.jpg')
    MAX_IMAGES = 10000

step 1

• find the first magic ball using (medianBlur -> HoughCircles)

def find_ball_with_opencv(path):
    img_rgb = cv2.imread(path)
    img = cv2.imread(path, cv2.IMREAD_GRAYSCALE)
    img_b = cv2.medianBlur(img,5)
    cimg = cv2.cvtColor(img_b,cv2.COLOR_GRAY2BGR)
    circles = cv2.HoughCircles(img_b,cv2.HOUGH_GRADIENT,1,300,param1=100,param2=30,minRadius=0,maxRadius=100)
    circles = np.uint16(np.around(circles))
    for c in circles[0,:]:
        cv2.circle(img_rgb,(c[0],c[1]),c[2],(0,255,0),7)
        cv2.circle(img_rgb,(c[0],c[1]),2,(0,0,255),7)
    bbox = [c[0]-c[2],c[1]-c[2],c[0]+c[2],c[1]+c[2]]
    return img_rgb,bbox

img_rgb,bbox = find_ball_with_opencv(config.magic_ball_for_example)
plt.imshow(img_rgb[:,:,::-1])
plt.show()

step 2

• Generate 10000 images with copy paste aug

transform = A.Compose([
    A.ShiftScaleRotate(p=1.0,shift_limit=0.1,scale_limit=0.25,rotate_limit=360,),
    A.Downscale(p=0.75),
    A.GaussianBlur(p=0.75)
])
def gen_random_magic_ball(color_image,transform,min_imgz=64):
    transformed = transform(image=color_image)
    transformed_image = transformed["image"]
    img=Image.fromarray(transformed_image)
    npImage=np.array(img)
    h,w=img.size
    alpha = Image.new('L', img.size,0)
    draw = ImageDraw.Draw(alpha)
    draw.pieslice([0,0,h,w],0,360,fill=255)
    npAlpha=np.array(alpha)
    
    if np.random.randint(2)==0:
        npImage=npImage.astype(float) * np.expand_dims(npAlpha.astype(bool).astype(float),-1)
        c = 0
    else:
        npImage=np.ones_like(npImage.astype(float)) * np.expand_dims(npAlpha.astype(bool).astype(float),-1)
        if np.random.randint(4) !=0:
            npImage[:,:,0] *= np.random.randint(256)
            npImage[:,:,1] *= np.random.randint(256)
            npImage[:,:,2] *= np.random.randint(256)
        elif np.random.randint(4) !=0:
            npImage *= np.random.randint(64) 
        else:
            npImage[:,:,1] *= np.random.randint(256)
            npImage[:,:,0] *= 0
            npImage[:,:,2] *= 0
        c=1
    size = min_imgz+np.random.randint(min_imgz)
    return cv2.resize(npImage.astype(np.uint8),(size,size)),c

def get_random_background(random_paths=config.random_paths):
    r_num = np.random.randint(len(random_paths))
    return cv2.imread(random_paths[r_num])

def copy_paste(back_image,img):
    h,w,_ = back_image.shape
    img_gen,c = gen_random_magic_ball(img,transform,min_imgz=min(h,w)//10)
    new_image = np.zeros_like(back_image)
    startY = np.random.randint(h-img_gen.shape[0])
    startX = np.random.randint(w-img_gen.shape[1])
    new_image[startY:startY+img_gen.shape[0],startX:startX+img_gen.shape[1]] = img_gen.copy()
    fg = cv2.bitwise_or(back_image,back_image, mask=(1-new_image.astype(bool)).astype(np.uint8)[:,:,0])
    fg2 = cv2.bitwise_or(new_image,new_image, mask=(new_image.astype(bool)).astype(np.uint8)[:,:,0])
    return fg+fg2,[startX,startY,startX+img_gen.shape[1],startY+img_gen.shape[0],c]

img_ball = cv2.imread(config.magic_ball_for_example)[:,:,::-1]
img_ball = img_ball[bbox[1]:bbox[3],bbox[0]:bbox[2]]
plt.imshow(img_ball)

<matplotlib.image.AxesImage at 0x7f662c1762d0>

plt.figure(figsize=[20,20])
for row in range(9):
    bboxes = []
    plt.subplot(3,3,row+1)
    results = get_random_background(config.random_paths)[:,:,::-1]
    for _ in range(1+ np.random.randint(10)):
        results,bbox2 = copy_paste(results,img_ball)
        bboxes.append(bbox2)
        if bbox2[-1]==0:
            c = 255
        else:
            c = 0
        cv2.rectangle(results,(bbox2[0],bbox2[1]),(bbox2[2],bbox2[3]),c,4)
    plt.imshow(results)

for num_image in tqdm(range(config.MAX_IMAGES)):
    bboxes = []
    results = get_random_background(random_paths)[:,:,::-1]
    for _ in range(1+ np.random.randint(5)):
        results,bbox2 = copy_paste(results,img_ball)
        bboxes.append(bbox2)
    cv2.imwrite(f"train/{num_image}.jpg",results[:,:,::-1])
    bboxes = np.array(bboxes).astype(float)
    h,w,_ = results.shape
    with open(f'train/{num_image}.txt', 'w+') as f:
        for b in bboxes:
            class_num = b[4]
            norm_w = (b[2]-b[0])/w
            norm_h = (b[3]-b[1])/h
            norm_x_center = b[0]/w + norm_w/2
            norm_y_center = b[1]/h + norm_h/2
            row = [class_num,norm_x_center,norm_y_center,norm_w,norm_h]
            row = np.array(row).astype(float).astype(str)
            text = ' '.join(row)
            f.write(text)
            f.write("\n")

all_data = glob.glob("train/*.jpg")
with open(f'train.txt', 'w+') as f:
    for row in all_data[:config.MAX_IMAGES-200]:
        f.write(row)
        f.write("\n")
with open(f'val.txt', 'w+') as f:
    for row in all_data[config.MAX_IMAGES-200:]:
        f.write(row)
        f.write("\n")

Train Yolov8 nano

• Train data with mixup and mosaic

from ultralytics import YOLO
model = YOLO("yolov8n.pt") 
results = model.train(data='ball.yaml', epochs=6, imgsz=640,save=True,exist_ok=True,plots=True,val=True,mixup=0.75,batch=32,close_mosaic=0,name="magic_ball")

Ultralytics YOLOv8.0.145 🚀 Python-3.7.10 torch-1.8.1+cu111 CUDA:0 (NVIDIA RTX 6000 Ada Generation, 48640MiB)
WARNING ⚠️ Upgrade to torch>=2.0.0 for deterministic training.
engine/trainer: task=detect, mode=train, model=yolov8n.pt, data=ball.yaml, epochs=6, patience=50, batch=32, imgsz=640, save=True, save_period=-1, cache=False, device=None, workers=8, project=None, name=magic_ball, exist_ok=True, pretrained=True, optimizer=auto, verbose=True, seed=0, deterministic=True, single_cls=False, rect=False, cos_lr=False, close_mosaic=0, resume=False, amp=True, fraction=1.0, profile=False, overlap_mask=True, mask_ratio=4, dropout=0.0, val=True, split=val, save_json=False, save_hybrid=False, conf=None, iou=0.7, max_det=300, half=False, dnn=False, plots=True, source=None, show=False, save_txt=False, save_conf=False, save_crop=False, show_labels=True, show_conf=True, vid_stride=1, line_width=None, visualize=False, augment=False, agnostic_nms=False, classes=None, retina_masks=False, boxes=True, format=torchscript, keras=False, optimize=False, int8=False, dynamic=False, simplify=False, opset=None, workspace=4, nms=False, lr0=0.01, lrf=0.01, momentum=0.937, weight_decay=0.0005, warmup_epochs=3.0, warmup_momentum=0.8, warmup_bias_lr=0.1, box=7.5, cls=0.5, dfl=1.5, pose=12.0, kobj=1.0, label_smoothing=0.0, nbs=64, hsv_h=0.015, hsv_s=0.7, hsv_v=0.4, degrees=0.0, translate=0.1, scale=0.5, shear=0.0, perspective=0.0, flipud=0.0, fliplr=0.5, mosaic=1.0, mixup=0.75, copy_paste=0.0, cfg=None, tracker=botsort.yaml, save_dir=runs/detect/magic_ball
Overriding model.yaml nc=80 with nc=2

                   from  n    params  module                                       arguments                     
  0                  -1  1       464  ultralytics.nn.modules.conv.Conv             [3, 16, 3, 2]                 
  1                  -1  1      4672  ultralytics.nn.modules.conv.Conv             [16, 32, 3, 2]                
  2                  -1  1      7360  ultralytics.nn.modules.block.C2f             [32, 32, 1, True]             
  3                  -1  1     18560  ultralytics.nn.modules.conv.Conv             [32, 64, 3, 2]                
  4                  -1  2     49664  ultralytics.nn.modules.block.C2f             [64, 64, 2, True]             
  5                  -1  1     73984  ultralytics.nn.modules.conv.Conv             [64, 128, 3, 2]               
  6                  -1  2    197632  ultralytics.nn.modules.block.C2f             [128, 128, 2, True]           
  7                  -1  1    295424  ultralytics.nn.modules.conv.Conv             [128, 256, 3, 2]              
  8                  -1  1    460288  ultralytics.nn.modules.block.C2f             [256, 256, 1, True]           
  9                  -1  1    164608  ultralytics.nn.modules.block.SPPF            [256, 256, 5]                 
 10                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 11             [-1, 6]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 12                  -1  1    148224  ultralytics.nn.modules.block.C2f             [384, 128, 1]                 
 13                  -1  1         0  torch.nn.modules.upsampling.Upsample         [None, 2, 'nearest']          
 14             [-1, 4]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 15                  -1  1     37248  ultralytics.nn.modules.block.C2f             [192, 64, 1]                  
 16                  -1  1     36992  ultralytics.nn.modules.conv.Conv             [64, 64, 3, 2]                
 17            [-1, 12]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 18                  -1  1    123648  ultralytics.nn.modules.block.C2f             [192, 128, 1]                 
 19                  -1  1    147712  ultralytics.nn.modules.conv.Conv             [128, 128, 3, 2]              
 20             [-1, 9]  1         0  ultralytics.nn.modules.conv.Concat           [1]                           
 21                  -1  1    493056  ultralytics.nn.modules.block.C2f             [384, 256, 1]                 
 22        [15, 18, 21]  1    751702  ultralytics.nn.modules.head.Detect           [2, [64, 128, 256]]           
Model summary: 225 layers, 3011238 parameters, 3011222 gradients, 8.2 GFLOPs

Transferred 319/355 items from pretrained weights
TensorBoard: Start with 'tensorboard --logdir runs/detect/magic_ball', view at http://localhost:6006/
AMP: running Automatic Mixed Precision (AMP) checks with YOLOv8n...
AMP: checks passed ✅
train: Scanning train... 9800 images, 0 backgrounds, 0 corrupt: 100%|██████████| 9800/9800 [00:09<00:00, 997.25it/s] 
train: New cache created: train.cache
albumentations: Blur(p=0.01, blur_limit=(3, 7)), MedianBlur(p=0.01, blur_limit=(3, 7)), ToGray(p=0.01), CLAHE(p=0.01, clip_limit=(1, 4.0), tile_grid_size=(8, 8))
/home/kaor/.conda/envs/mmdet/lib/python3.7/site-packages/torch/utils/data/dataloader.py:477: UserWarning: This DataLoader will create 8 worker processes in total. Our suggested max number of worker in current system is 6, which is smaller than what this DataLoader is going to create. Please be aware that excessive worker creation might get DataLoader running slow or even freeze, lower the worker number to avoid potential slowness/freeze if necessary.
  cpuset_checked))
val: Scanning train... 200 images, 0 backgrounds, 0 corrupt: 100%|██████████| 200/200 [00:00<00:00, 1213.90it/s]
val: New cache created: train.cache
/home/kaor/.conda/envs/mmdet/lib/python3.7/site-packages/torch/utils/data/dataloader.py:477: UserWarning: This DataLoader will create 16 worker processes in total. Our suggested max number of worker in current system is 6, which is smaller than what this DataLoader is going to create. Please be aware that excessive worker creation might get DataLoader running slow or even freeze, lower the worker number to avoid potential slowness/freeze if necessary.
  cpuset_checked))
Plotting labels to runs/detect/magic_ball/labels.jpg... 
optimizer: AdamW(lr=0.001667, momentum=0.9) with parameter groups 57 weight(decay=0.0), 64 weight(decay=0.0005), 63 bias(decay=0.0)
Image sizes 640 train, 640 val
Using 8 dataloader workers
Logging results to runs/detect/magic_ball
Starting training for 6 epochs...

      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size
        1/6      5.93G     0.4592     0.8717     0.8664         77        640: 100%|██████████| 307/307 [00:55<00:00,  5.56it/s]
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:01<00:00,  3.12it/s]
                   all        200        575      0.999      0.965       0.97      0.954

      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size
        2/6      5.96G     0.3396     0.3815     0.8254         83        640: 100%|██████████| 307/307 [00:53<00:00,  5.72it/s]
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:01<00:00,  3.06it/s]
                   all        200        575      0.998      0.969      0.972      0.965

      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size
        3/6      5.96G     0.3104     0.3271     0.8191         88        640: 100%|██████████| 307/307 [00:52<00:00,  5.87it/s]
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:01<00:00,  3.23it/s]
                   all        200        575      0.997      0.966      0.971      0.965

      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size
        4/6      5.94G     0.2918     0.2997      0.815         57        640: 100%|██████████| 307/307 [00:52<00:00,  5.86it/s]
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:01<00:00,  3.12it/s]
                   all        200        575          1       0.97      0.971      0.968

      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size
        5/6      5.93G     0.2813     0.2852     0.8131         77        640: 100%|██████████| 307/307 [00:52<00:00,  5.83it/s]
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:01<00:00,  2.95it/s]
                   all        200        575          1       0.97      0.972      0.969

      Epoch    GPU_mem   box_loss   cls_loss   dfl_loss  Instances       Size
        6/6      5.96G     0.2658     0.2674     0.8104         52        640: 100%|██████████| 307/307 [00:52<00:00,  5.81it/s]
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:02<00:00,  1.39it/s]
                   all        200        575          1       0.97      0.974      0.971

6 epochs completed in 0.094 hours.
Optimizer stripped from runs/detect/magic_ball/weights/last.pt, 6.2MB
Optimizer stripped from runs/detect/magic_ball/weights/best.pt, 6.2MB

Validating runs/detect/magic_ball/weights/best.pt...
Ultralytics YOLOv8.0.145 🚀 Python-3.7.10 torch-1.8.1+cu111 CUDA:0 (NVIDIA RTX 6000 Ada Generation, 48640MiB)
Model summary (fused): 168 layers, 3006038 parameters, 0 gradients, 8.1 GFLOPs
                 Class     Images  Instances      Box(P          R      mAP50  mAP50-95): 100%|██████████| 4/4 [00:02<00:00,  1.55it/s]
                   all        200        575          1       0.97      0.974      0.971
            magic_ball        200        293          1      0.997      0.995      0.993
                  ball        200        282          1      0.943      0.952       0.95
Speed: 1.1ms preprocess, 0.5ms inference, 0.0ms loss, 0.5ms postprocess per image
Results saved to runs/detect/magic_ball

Yolo results

• I trained only 6 epochs because I haven't really validated the datasets, and I try to avoid overfitting.

Evaluation

• Using 2Xscale TTA

start_time = time.time()
all_images = []
paths = glob.glob(f"{path}*")
for p in paths:
    s = time.time()
    img = cv2.imread(p)
    h,w,_ = img.shape
    n = max(h,w)+1
    res640 = model(img[:,:,::-1],
                save=False,
                augment=True,
                imgsz=640,
                conf=0.3,
                classes=[0,1],
                exist_ok=True)
    res320 = model(img[:,:,::-1],
                save=False,
                augment=True,
                imgsz=320,
                conf=0.3,
                classes=[0,1],
                exist_ok=True)
    res320 = res320[0].boxes.data.cpu().numpy()/n
    res640 = res640[0].boxes.data.cpu().numpy()/n
    res320 = res320[res320[:, 5].argsort()]
    res640 = res640[res640[:, 5].argsort()]
    if len(res320)==0 and len(res640)!=0:
        boxes_list = [res640[:,:4].tolist()]
        scores_list = [res640[:,4].tolist()]
        labels_list = [res640[:,5].tolist()]
    elif len(res640)==0 and len(res320)!=0:
        boxes_list = [res320[:,:4].tolist()]
        scores_list = [res320[:,4].tolist()]
        labels_list = [res320[:,5].tolist()]
    elif len(res640)!=0 and len(res320)!=0:    
        boxes_list = [res320[:,:4].tolist(),res640[:,:4].tolist()]
        scores_list = [res320[:,4].tolist(),res640[:,4].tolist()]
        labels_list = [res320[:,5].tolist(),res640[:,5].tolist()]
    else:
        all_images.append([p,img,time.time()-s])
        continue
    boxes, scores, labels = nms(boxes_list, scores_list, labels_list, weights=[1.0,1.0], iou_thr=0.1)
    boxes*=n
    wx = boxes[:,2] - boxes[:,0]
    hy = boxes[:,3] - boxes[:,1]
    xcenter = boxes[:,0] + wx/2
    ycenter = boxes[:,1] + hy/2
    for x,y,ww,hh,l in zip(xcenter,ycenter,wx,hy,labels):
        if ww < 20 or hh<20:
            continue
        raduis = max(ww,hh)
        if l==0:
            cv2.circle(img,(int(x),int(y)),int(raduis/2.0),(200,50,100) if l==0 else (0,255,0),5)
        # else:
        #     cv2.circle(img,(int(x),int(y)),int(5),(255,0,0) if l==0 else (0,255,0),10)
    all_images.append([p,img,time.time()-s])
all_time_p = time.time()-start_time
clear_output()
print(f"cell time: {all_time_p}")

cell time: 0.9948115348815918

plt.figure(figsize=[20,20])
for i,(p,im,t) in enumerate(all_images):
    plt.subplot(4,4,i+1)
    plt.imshow(im[:,:,::-1])
    plt.title(f"predict time - {round(t*1000,3)} ms")
    cv2.imwrite(p.replace("Balls","results"),im)

Conclusions

• According to the plot above, you can see that all the magic balls were found, and no other ball was identified as a magic ball.
• Apart from one image (MVC-008F.JPG), the bounding box around the magic ball is pretty accurate.

• The model handles a manipulated image well (bonus).

• The average processing time for one image is 70 milliseconds, while the whole test set of images took 1.12 seconds to process.